search

kenshohara / 3D-ResNets-PyTorch

3D ResNets for Action Recognition (CVPR 2018)
MIT License
3.9k stars 932 forks source link

Very poor performance on pre-trained models.. maybe un matching Labels to index file #243

Closed Purav-Zumkhawala closed 3 years ago

Purav-Zumkhawala commented 3 years ago

Hi, I am trying to test the r3d50_K_200ep.pth model on youtube videos and the pre-trained model is predicting very poorly.. I am not getting any correct predictions which make me wonder that the class list mapping to index that I have, is not the same one that was used while training. I am using the kinetics-700 class list in sorted order. The command I am using is the following for generating the inferences

"python main.py --root_path "+root_path+" --video_path mini_clips_extracted_jpg\\Shaving --annotation_path test.json "+"--result_path results --dataset kinetics --resume_path "+model_path+" --model_depth 50 --n_classes 700 --no_train --no_val --inference --output_topk 1 --inference_batch_size 1"

I have used the same command for predicting on the UCF101 dataset after fine-tuning it on the kinetic-400 pre-trained model and it performs really well like given in the papers.

Does anyone have any idea what might be causing these issues. @kenshohara @guilhermesurek

Purav-Zumkhawala commented 3 years ago

The issue was as I suspected the class labels while training and the ones that I was using the sorted labels were not matching and thus I was getting low accuracy.

Look at this issue for the labels that were used for training. The answer is given by @Jimmy880. Issue link: #211